Review of fall risk assessment in geriatric populations using inertial sensors

A literature review was performed using the key words: “(aged or geriatr* or gerontol* or senior* or elder* or old*) and (acceler* or inertia* or gyro*) and (fall* or fall risk or fall prediction)” on Web of Science, Scholars Portal, Pubmed, and Google Scholar on March 25 2013. Reference lists from the identified publications were reviewed to identify additional research articles of interest. The results of this search are shown in Figure 1. Paper inclusion criteria included fall risk assessment using inertial sensors, involvement of a geriatric population based on a mean participant age greater than or equal to 60 years, and published in English.

Forty articles from the database searches and article references met the inclusion criteria. Pertinent methodological features were extracted; including, fall risk classification method, sensor placement, activities assessed while inertial sensors were worn, and inertial-based parameters used to assess fall risk. Model assessment outcomes (i.e. accuracy, specificity, and sensitivity [28]) were extracted when available.

Three main methods were used to classify subjects into faller and non-faller categories: retrospective fall history (30%), prospective fall occurrence (15%), and scores on clinical assessments (32.5%). A combination of retrospective fall history and clinical assessment tools were used to assess fall risk in 22.5% of the studies. Brief descriptions of the clinical assessment tools are provided in Table 1. Table 2 lists fall risk classification methods used in the literature.

Two different inertial sensors were used in the reviewed papers: gyroscopes (angular velocity) and accelerometers (linear acceleration) [76]. Descriptions of gyroscopes and accelerometers are provided by Webster [76]. Accelerometers were the only inertial sensor in 70% of the studies, whereas gyroscopes were the only inertial sensor in one study [17]. Both accelerometers and gyroscopes were used in 27.5% of the studies.

Accelerometers and gyroscopes are small enough for measurement during activity, via attachment to a body part, belt, or headband. The lower back, including the pelvis, sacrum, and the L3 to L5 vertebrae, is the most common sensor location and was the only location in 65% of the studies. This site approximates the center of mass location [46, 49, 60, 61, 74] and is acceptable for long-term at-home use [49, 61]. Other sensor locations include the head [18, 55, 64], upper back [44, 73, 75], sternum [17, 42, 45], shoulder [41], elbow [41], wrist [41], hip [18, 56], thigh [42], knee [41, 56], shank [50], ankle [41, 56], and foot [69, 71].

Various activities were used for inertial-sensor-based fall risk assessment. The most frequently assessed activity was level ground walking (45%), followed by Timed Up and Go (TUG) (32.5%), sit-to-stand transitions (STS, 22.5%), standing postural sway (20%), left-right Alternating Step Test (AST) on level ground (15%), and uneven-ground walking (2.5%). Many studies used a combination of activities (20%). For level walking, subject-selected walking speed was assessed in the majority of studies (66.7%) while other walking speeds (slow, fast) were assessed in 26.7% of studies and a designated speed was assessed in one study [73].

One hundred and thirty distinct variables were assessed in the literature, and can be categorized as: position and angle (7.7%), angular velocity (11.5%), linear acceleration (20%), spatial (3.8%), temporal (23.1%), energy (3.8%), frequency (15.4%), and other (14.6%). All variables that had significant outcomes (p < 0.05), together with sensor body locations, are presented in Table 3.

Of the variables that were assessed in more than one study, only 13 were significant (p < 0.05) each time they were assessed: 1) mediolateral and anteroposterior postural sway length [43, 51]; 2) mediolateral and anteroposterior postural sway velocity [43, 51]; 3) ratio of mean squared modulus for postural sway [46‐48]; 4) standard deviation of anteroposterior acceleration [18, 74]; 5) root mean square amplitude of vertical linear acceleration [55, 72]; 6) gait speed [40, 44, 55, 64, 72, 74]; 7) sit-to-stand transition duration [17, 45]; 8) dominant Fast Fourier Transform (FFT) peak parameters derived from lower-back linear acceleration signals [59, 61, 62]; 9) ratio of even to odd harmonic magnitudes derived from head, upper back, and lower-back linear acceleration signals [44, 55, 57, 58, 64, 72, 75]; 10) area under the first six harmonics divided by the remaining area for lower-back linear acceleration signals [57, 58]; 11) ratio of the first four harmonics to the magnitude of the first six harmonics for lower-back linear acceleration signals [57, 58]; 12) maximum Lyapunov exponent of angular velocity signal [73, 77]; 13) discrete wavelet transform parameters from lower-back angular velocity and linear acceleration signals and sternum linear acceleration signals [45, 63]. Six of these multi-study variables (1,2,5,6,9,12) were from different research groups, while seven variables (3,4,7,8,10, 11,13) were from a single research group.

Mediolateral and anteroposterior postural sway length and velocity are measures of postural stability that represent trunk movement during static standing [43, 51]. The root mean square of vertical linear accelerations has been used to measure gait smoothness, with larger values linked to increased fall risk [55, 72]. Low gait speed has been identified in the broader literature as an indicator of fall risk [77, 78]; however, adoption of a low gait speed may also be an accommodation linked to fear of falling and not an indicator of high risk of falling [79]. The ratio of even to odd harmonics of linear acceleration reflects the proportion of the acceleration signal that is in phase with the participant’s stride frequency, with even harmonics correlating with in-phase components and odd harmonics correlating with out-of-phase components [55, 75]. The area under the first six harmonics divided by the remaining area under the magnitude spectrum curve of the lower-back linear accelerations provides a quantification of gait periodicity [57, 58], and the ratio of the first four harmonics to the magnitude of the first six harmonics for lower-back linear accelerations indicates the pattern of harmonic magnitudes [58]. The maximum Lyapunov exponent of angular velocity reflects local dynamic stability [56, 73] and converts time series measurements into a state space [73]. In the state space, the Lyapunov exponents measure the average rate of expansion or contraction of the original trajectory in response to perturbations. Larger Lyapunov exponent values indicate a decreased ability to compensate for local perturbations during gait, increased instability, and increased fall risk [56, 73].

A wide range of variables have been incorporated into fall risk classification models. Variables that achieved high fall risk classification sensitivity and specificity (i.e., greater than or equal to 80%) are discussed here. Successful single-variable models used: 14) the mean squared modulus ratio for postural sway derived from lower back angular velocity [46‐48], 15) sit-to-stand transition duration determined from sternum acceleration and angular velocity [45], 16) sit-to-stand fractal dimension derived from sternum acceleration and angular velocity [45], and 17) sit-to-stand lower back jerk (derivative of acceleration) [74]. The mean squared modulus ratio for postural sway is the ratio of the rotational kinematic energy during eyes-open or closed postural sway on a foam surface, to the rotational kinematic energy during eyes-open postural sway on a firm surface [46‐48]. The sit-to-stand fractal dimension represents the regularity of the sit-to-stand movement, with larger values associated with greater movement irregularity and fall risk [45].

Several multi-variable models also achieved sensitivity and specificity levels greater than or equal to 80%. One model included 18) pelvic sway and 19) kinetic energy [49]. Caby et al., 2011 [41] investigated a variety of single-variable and multi-variable models derived from accelerometers on the knee, ankle, shoulder, elbow, and wrist. These variables included correlation between left and right elbow accelerations, step frequency, jerk, spectral entropy, acceleration frequency parameters, and 25 m walk time normalized to participant height. Liu et al., 2011 [57] developed a multi-variable model that included 126 temporal, energy, and frequency variables derived from TUG, STS, and AST movements.

In half of the papers, geriatric fall risk was predicted using derived models (50%), as opposed to correlating a variable with fall risk or fall occurrence. Regression models (65%), mathematical classifiers (25%), decision trees (15%), neural networks (15%), support vector machines (10%), and cluster analysis (10%) were employed to predict fall risk, with some studies using more than one method (30%). The accuracy, specificity, and sensitivity [28] of these models are shown in Table 4.

A wide range of sensitivity (55-100%) and specificity (15-100%) levels have been reported for inertial-sensor-based fall risk assessment models (Table 4). The highest sensitivity and specificity levels for retrospective fall history were 91.3% and 83.3%, respectively, and the highest levels using prospective fall history were 74% and 82%, respectively. However, 50% of the studies did not employ separate data sets for model training and validation, thereby limiting the model’s applicability beyond the training set population. Validating the model with the training population would likely result in inflated accuracy, specificity, and sensitivity levels when the model is applied to the general population compared to model validation with a different testing population. The five best performing models, in terms of overall accuracy, specificity, and sensitivity, used neural networks [47, 48], naive Bayesian classifier [41], Mahalanobis cluster analysis [46], and a decision tree [49]. The five worst performing models used regression [44, 53, 62, 72] and a support vector machine [51]. Therefore, intelligent computing methods (neural networks, Bayesian classifiers, etc.) may be more appropriate for fall risk classification than regression techniques.

In the majority of papers, sensors were placed on the lower back. However, the justification for this location is limited to an intention to approximate the body center of mass and the location’s unobtrusive nature for long-term use. High subject acceptance for long-term lower back sensor placement was found in a 20 day case-study by Giansanti et al., 2009 [80]. While the lower back is a promising sensor location, the upper back, hip, and thigh have potential as long-term sensor locations [81]. To date, there has been no objective evaluation to determine which sensor sites, or combinations of sites, provide the most reliable fall risk assessment. Other sensor locations, along the legs and arms, were used in predictive models with comparable sensitivity and specificity [41, 50]. Current research has not confirmed if the total body center of gravity region is superior to other locations for fall risk identification.

Of the 13 variables that were assessed in more than one study and had significant outcomes in each study (p < 0.05) and the six variables that had high fall risk classification sensitivity and specificity, 58% were related to postural instability and gait consistency over strides (1–3,9-14,16,18). Further research is needed to corroborate these initial findings toward identifying a set of inertial-sensor-based variables that yield a robust and accurate fall risk assessment model and clinical tool.

Clinical fall risk assessment was the predominate criterion method for classifying fallers and non-fallers when evaluating sensor-derived fall risk classification, with 32.5% of the studies using clinical assessment as the only means of classification. However, clinical assessments include false positives and false negatives that introduce inaccuracies when evaluating sensor-based systems. Another concerning aspect is that fall risk thresholds were not used consistently across research studies. For example, three studies that assessed geriatric populations devoid of neurological disorders used different TUG completion-time thresholds to determine fall risk, ranging from 11 s to 20 s [40, 52, 59]. The literature recommends a threshold of 14 s for studies focused on community dwelling elderly without neurological disorders [37] and 30 s for community dwelling elderly with neurological disorders [82].

Retrospective fall assessment was the only criterion classification method in 30% of the studies; however, a fall would already have occurred before the study assessment. The gait strategies and patterns during the fall may have been different from those during the assessment, because the participants could have adjusted their walking and mobility patterns to be more conservative, stable, and safe. Prospective fall occurrence is the preferred criterion classification method when placing subjects into low risk (non-faller) and high risk (faller) categories, but prospective risk was only used in 15% of studies. Since the goal of fall risk assessment is to predict the likelihood of future falls, a prospective study that uses fall occurrence records after assessment would be more appropriate. Furthermore, prospective fall assessments have greater accuracy compared to retrospective fall assessment due to patient fall recollection issues with retrospective assessment [83].

A thorough assessment of optimal inertial sensor sites should be performed in conjunction with identification of optimal inertial-sensor-based variables, since the usefulness of a variable will likely be site specific. Better knowledge of user compliance with sensor placement at different sites will also be important when developing assessment tools that are widely accepted in the clinical and community environments. A generalized focus-group assessment of user acceptance of wearable sensors for fall detection determined that wearable sensors would be accepted provided they were unobtrusive [84]. Only one case study demonstrated the acceptability of the lower back sensor location [80]. Therefore, larger, long-term, user compliance studies of wearable sensors are needed to ensure feasibility for fall risk assessment. If long-term compliance is an issue, sensor use may be confined to short periods in a clinical environment.

Future studies should prospectively assess fall risk instead of relying on clinical assessments or a retrospective fall assessment, as previously discussed. Several papers emphasized the importance of using prospective fall risk occurrence in their future research [51, 54, 60, 62, 72‐74], thereby avoiding the two biggest limitations of retrospective fall assessment: inaccurate recollection of fall history and gait changes due to past falls.

An important next step in this research field is to expand assessments from the geriatric population to include more specialized populations; including, those with Alzheimer’s and Parkinson’s diseases. Inertial sensors have the potential to identify patients with these diseases in the early stages, as suggested by Giansanti et al., 2008 [47], and identify those at risk of falling. One paper identified fallers from non-fallers in a Parkinson’s Disease elderly subgroup [55]. However, more disease-specific research is needed since optimal sensor site(s) and sensor features may be different from the generalized geriatric population. For example, arm swing may be a critical feature for people with Parkinson’s Disease because reduced arm swing is one of the first presenting gait impairment features [85].

Another critical area for future work is to match predictive variables with specific fall risk factors [58]. These fall risk factors could include physical factors such as peripheral neuropathy, muscle weakness, visual impairments, reduced flexibility, and lower limb arthritis. This is a crucial step that would increase the clinical value of inertial-sensor-based fall risk assessment tools and allow identification of specific impairments that increase fall risk, thus enabling individualized treatment and intervention. The walking environment (i.e. level ground, stairs, ramps) and the activity performed (i.e. transition from sitting to standing, walking outdoors) could be associated with increased fall risk. Identification of such high-risk environments and activities may increase the accuracy of fall risk assessment [86]. Identification of specific risk factors is clinically important for preventing falls and their sequelae [79].

To demonstrate the utility of inertial-sensor-based fall risk assessments, a comparison between inertial-sensor-based assessments and current clinical assessments must be made in relation to prospective fall occurrences. This would determine whether inertial-sensor-based fall risk assessment methods alone can provide better accuracy than current clinical assessments, whether a combination of inertial sensors and current clinical assessment would optimize fall risk prediction, and if specific risk factors can be better identified when using inertial-sensor-based information. While some work has been done comparing inertial-based assessment to clinical assessments [1, 50], and using a combination of inertial and clinical assessments [60, 61], more research in this area is required to conclusively demonstrate the advantages of inertial-based fall risk assessments.